Which Sume video models accept a music track as an audio reference?

Per the Sume videos docs, the Seedance 2.x models, Wan 3.0, MiniMax H3 and MiniMax H3 Max accept audio references, and Gemini Omni Flash 1.1, higgsfield-genjutsu and h3-max-recast accept video references but not audio. If you want to pass a generated Lyria track into the video request itself, choose from the first group. For Omni, the bed has to be added afterwards in a Timeline render.

The Omni page adds that there is no reference_audio_urls field and that native audio is always on, so an Omni clip arrives with its own sound. The table prices an 8-second clip from each model with a $0.125 bed, so you can see what the choice costs. Prices change, so treat the arithmetic as a template and re-read the catalog before budgeting a campaign.

The table

Per-second prices are Sume list prices (default options, read 2026-10-08). The Omni row needs a $0.10 Timeline render to mix a bed in, since it cannot take one as a reference. The other rows show footage plus the bed, with no render, because the bed travels in the request.

What a reference does and does not do

An audio reference is an input the model may use. The docs list it under supported_input_references, but they do not say the output reproduces the track note for note, and they do not promise sync to your bed. Listen to the result, as the Music 1.0 page tells you to do for generated audio.

If you need the exact Lyria file under the picture, the reliable route is a Timeline render, which takes the track as a spine or as a soundtrack bed. It costs $0.10 per started minute. Passing the reference is the cheaper-to-describe route; the render is the exact one.

Choosing

For a lip-synced or beat-matched shot, test a reference on the cheapest model in the group first. An 8-second MiniMax H3 clip at 768p is $0.60, and the bed is $0.125. If the result is not what you wanted, you have spent $0.725 and can move to a render with the exact file.

Check the live envelope before you build: GET /v1/video-router/models returns supported_input_references and generate_audio per model, and the catalog can change. A short test request that sets only the fields you need will show quickly whether the model accepts your audio URL, and a refused field fails early instead of after the footage is billed. The Videos page shows the same fields.